refactor(factorminer): 禁用 npz 信号缓存并将库 I/O 对接本地 DSL

- 为 Factor 数据类新增 metadata 字段，用于标记未实现算子（unsupported）
- save_library 废弃 save_signals 参数，内部强制忽略，仅持久化 JSON 元数据，不再写入 .npz
- load_library 删除 .npz 恢复逻辑；加载时自动将 # TODO 公式的 unsupported 标记设为 True
- import_from_paper() 直接基于已本地化的 PAPER_FACTORS 构建库，并同步标记 TODO 公式
- 新增 tests/test_factorminer_library_io.py，覆盖序列化、加载及 paper factors 导入

src/factorminer/core/factor_library.py

@@ -39,6 +39,7 @@ class Factor:
     signals: Optional[np.ndarray] = field(default=None, repr=False)  # (M, T)
     research_metrics: dict = field(default_factory=dict)
     provenance: dict = field(default_factory=dict)
+    metadata: dict = field(default_factory=dict)
 
     def __post_init__(self) -> None:
         if not self.admission_date:
@@ -59,6 +60,7 @@ class Factor:
             "admission_date": self.admission_date,
             "research_metrics": self.research_metrics,
             "provenance": self.provenance,
+            "metadata": self.metadata,
         }
 
     @classmethod
@@ -77,6 +79,7 @@ class Factor:
             admission_date=d.get("admission_date", ""),
             research_metrics=d.get("research_metrics", {}),
             provenance=d.get("provenance", {}),
+            metadata=d.get("metadata", {}),
         )
 
 
@@ -172,7 +175,7 @@ class FactorLibrary:
             # Pearson on ranks == Spearman
             ra_c = ra - ra.mean()
             rb_c = rb - rb.mean()
-            denom = np.sqrt((ra_c ** 2).sum() * (rb_c ** 2).sum())
+            denom = np.sqrt((ra_c**2).sum() * (rb_c**2).sum())
             if denom < 1e-12:
                 continue
             corr_sum += abs((ra_c * rb_c).sum() / denom)
@@ -206,9 +209,7 @@ class FactorLibrary:
         (admitted, reason) : Tuple[bool, str]
         """
         if candidate_ic < self.ic_threshold:
-            return False, (
-                f"IC {candidate_ic:.4f} below threshold {self.ic_threshold}"
-            )
+            return False, (f"IC {candidate_ic:.4f} below threshold {self.ic_threshold}")
 
         if self.size == 0:
             return True, "First factor in library"
@@ -221,9 +222,7 @@ class FactorLibrary:
                 f"{self.correlation_threshold} with existing library factor"
             )
 
-        return True, (
-            f"Admitted: IC={candidate_ic:.4f}, max_corr={max_corr:.4f}"
-        )
+        return True, (f"Admitted: IC={candidate_ic:.4f}, max_corr={max_corr:.4f}")
 
     def check_replacement(
         self,
@@ -258,8 +257,10 @@ class FactorLibrary:
         (should_replace, factor_to_replace_id, reason) : Tuple[bool, Optional[int], str]
         """
         if candidate_ic < ic_min:
-            return False, None, (
-                f"IC {candidate_ic:.4f} below replacement floor {ic_min}"
+            return (
+                False,
+                None,
+                (f"IC {candidate_ic:.4f} below replacement floor {ic_min}"),
             )
 
         if self.size == 0:
@@ -277,21 +278,33 @@ class FactorLibrary:
                 correlated_factors.append((fid, corr, factor.ic_mean))
 
         if len(correlated_factors) != 1:
-            return False, None, (
-                f"Found {len(correlated_factors)} correlated factors "
-                f"(need exactly 1 for replacement)"
+            return (
+                False,
+                None,
+                (
+                    f"Found {len(correlated_factors)} correlated factors "
+                    f"(need exactly 1 for replacement)"
+                ),
             )
 
         fid, corr, existing_ic = correlated_factors[0]
         if candidate_ic < ic_ratio * existing_ic:
-            return False, None, (
-                f"IC {candidate_ic:.4f} < {ic_ratio} * {existing_ic:.4f} = "
-                f"{ic_ratio * existing_ic:.4f}"
+            return (
+                False,
+                None,
+                (
+                    f"IC {candidate_ic:.4f} < {ic_ratio} * {existing_ic:.4f} = "
+                    f"{ic_ratio * existing_ic:.4f}"
+                ),
             )
 
-        return True, fid, (
-            f"Replace factor {fid}: candidate IC {candidate_ic:.4f} > "
-            f"{ic_ratio} * {existing_ic:.4f}, corr={corr:.4f}"
+        return (
+            True,
+            fid,
+            (
+                f"Replace factor {fid}: candidate IC {candidate_ic:.4f} > "
+                f"{ic_ratio} * {existing_ic:.4f}, corr={corr:.4f}"
+            ),
         )
 
     # ------------------------------------------------------------------
@@ -321,8 +334,11 @@ class FactorLibrary:
 
         logger.info(
             "Admitted factor %d '%s' (IC=%.4f, max_corr=%.4f, category=%s)",
-            factor.id, factor.name, factor.ic_mean,
-            factor.max_correlation, factor.category,
+            factor.id,
+            factor.name,
+            factor.ic_mean,
+            factor.max_correlation,
+            factor.category,
         )
         return factor.id
 
@@ -360,7 +376,10 @@ class FactorLibrary:
 
         logger.info(
             "Replaced factor %d with %d '%s' (IC=%.4f)",
-            old_id, new_factor.id, new_factor.name, new_factor.ic_mean,
+            old_id,
+            new_factor.id,
+            new_factor.name,
+            new_factor.ic_mean,
         )
 
     def remove_factor(self, factor_id: int) -> None:
@@ -381,9 +400,7 @@ class FactorLibrary:
     # Correlation matrix management
     # ------------------------------------------------------------------
 
-    def _max_correlation_with_library(
-        self, candidate_signals: np.ndarray
-    ) -> float:
+    def _max_correlation_with_library(self, candidate_signals: np.ndarray) -> float:
         """Compute max |rho| between candidate and all library factors."""
         max_corr = 0.0
         for factor in self.factors.values():
@@ -453,9 +470,7 @@ class FactorLibrary:
                 self.correlation_matrix[idx, other_idx] = 0.0
                 self.correlation_matrix[other_idx, idx] = 0.0
                 continue
-            corr = self._compute_correlation_vectorized(
-                factor.signals, other.signals
-            )
+            corr = self._compute_correlation_vectorized(factor.signals, other.signals)
             self.correlation_matrix[idx, other_idx] = corr
             self.correlation_matrix[other_idx, idx] = corr
 
@@ -509,10 +524,7 @@ class FactorLibrary:
 
     def get_factors_by_category(self, category: str) -> List[Factor]:
         """Return all factors matching a given category."""
-        return [
-            f for f in self.factors.values()
-            if f.category == category
-        ]
+        return [f for f in self.factors.values() if f.category == category]
 
     def get_diagnostics(self) -> dict:
         """Library diagnostics: avg |rho|, max tail correlations, per-category counts, saturation.
@@ -539,8 +551,7 @@ class FactorLibrary:
 
         diag["category_counts"] = dict(cat_counts)
         diag["category_avg_ic"] = {
-            cat: cat_ic_sums[cat] / cat_counts[cat]
-            for cat in cat_counts
+            cat: cat_ic_sums[cat] / cat_counts[cat] for cat in cat_counts
         }
 
         # Correlation statistics
@@ -575,9 +586,7 @@ class FactorLibrary:
         Returns a lightweight dictionary suitable for inclusion in LLM prompts
         or memory store entries.
         """
-        factors_sorted = sorted(
-            self.factors.values(), key=lambda f: f.id, reverse=True
-        )
+        factors_sorted = sorted(self.factors.values(), key=lambda f: f.id, reverse=True)
         recent = factors_sorted[:5]  # Last 5 admissions
 
         categories = defaultdict(int)